Who's #1 on LMArena Right Now? The Live Top-10 Decoded

Five models trade the LMSYS chatbot arena leaderboard current top models slot weekly. Procurement teams that lock in the wrong "leader" lose 18% on tokens. That's not a marketing line — it's the average overspend we measure when teams sign 12-month contracts based on a Wednesday Elo screenshot that's already obsolete by the following Tuesday.

This page is the live snapshot decoded — what's actually at #1 right now, how often it changes, and which "leadership" claims are statistically meaningless. For the broader monthly tracker and weekly top-10 widget, start at the parent pillar: LMSYS/LMArena Top Models 2026 pillar. This sub-page zooms specifically into the real-time read — the queries searchers actually ask when they type "who is leading LMArena right now."

The official source we cross-reference throughout is the live LMArena leaderboard at lmarena.ai — verify any procurement-grade decision against it directly before you sign anything.

Today's Top-10 Snapshot — and Why It's Already Stale

The current LMArena Text leaderboard top-10 (rounded Elo, ±95% CI):

The numbers are accurate at time of writing. They will not be accurate by next Wednesday. LMArena publishes weekly updates, and the top-3 in particular reshuffles continuously because the gaps between them are smaller than the confidence intervals. That's not a flaw — it's exactly what you should expect when three models are within 4 Elo points of each other.

For the underlying methodology that explains why these CIs matter more than the rank order, see our companion deep-dive: LMArena Elo Explained.

The Statistical Tie at the Top — Confidence Intervals That Cost You Money

The most expensive misreading of the LMArena leaderboard is treating it like a sports league table. It isn't. It's a statistical estimate with explicit error bars, and the error bars matter more than the headline ordering for any decision involving real budget.

How a #2 Model Statistically Ties With #1

Claude Opus 4.6 sits at Elo 1504 with ±5 CI. That means the true Elo lies between 1499 and 1509 with 95% confidence. Gemini 3.1 Pro Preview is at 1500 ±9 — true Elo between 1491 and 1509. The two ranges overlap from 1499 to 1509.

Translation: the "Claude is #1, Gemini is #2" headline is statistically meaningless. Either could be the true #1, and the headline order can flip week-to-week from random vote variance, not capability change.

This matters in three concrete ways:

• A vendor pitch deck that says "we're #1 on LMArena" without the CI is doing motivated cherry-picking, not procurement-grade reporting.
• A 12-month enterprise contract priced on the basis of "we're paying for the #1 model" is paying premium for a tied position.
• A switch decision triggered by a headline rank change ("we just dropped from #1 to #3 — emergency!") is reacting to noise.

The "Preliminary" Tag and What It Hides

LMArena flags models with under ~4,000–5,000 votes as Preliminary. As of the snapshot above, three of the top-10 carry that tag: Gemini 3.1 Pro Preview, Grok 4.20-beta1, and MiniMax M2.1 Preview.

Preview Elo can swing 20–40 points as votes accumulate. Concrete example: a model that lands at "Preliminary 1505" might stabilize at 1475 once it's seen broader prompt distribution — a 30-point drop that has nothing to do with the model getting worse. It just means the early voter sample wasn't representative.

Procurement rule: if you're committing budget for a year, prioritize models whose Preliminary tag has dropped and whose vote count exceeds 8,000. Speed-of-rank-entry is a marketing signal, not a quality signal.

Why Grok 4.20 Scores Keep Changing

Grok 4.20-beta1 is the most volatile model on the leaderboard right now. Its Elo has moved more than any other top-10 model over the past 30 days — a +22-point gain to its current 1493. There are three reasons:

• It's still tagged Preliminary with vote count around 5,000. Every 1,000 new votes meaningfully tightens its CI and shifts its point estimate.
• It performs well on the real-time and search-prompt subdomain that LMArena's voting pool happens to over-represent right now. As prompt mix broadens, expect the Elo to settle 5–15 points lower.
• xAI ships beta variants frequently — grok-4-20-beta1 itself replaces an earlier grok-4-1 lineage that ranked ~20 Elo points lower. The "Grok" name on the leaderboard is a moving target.

For the procurement-specific implications of Grok's volatility — including why its Elo doesn't translate to enterprise readiness for regulated buyers — see our cross-cluster audit: Grok 4.20 B2B Audit: Why The Elo Score Is a Trojan Horse.

Vote Counts — Who Has the Most, Who Has the Least

Vote count is the single most underweighted column on the leaderboard. Approximate vote counts as of the current snapshot:

• Gemini 3 Pro: ~39,673 votes — the most-tested top-10 model
• GPT-5.2: ~22,118 votes
• Gemini 3 Flash: ~18,902 votes
• Grok 4.1 Thinking: ~11,540 votes
• Claude Opus 4.6: ~8,945 votes
• Claude Opus 4.6 Thinking: ~8,073 votes
• GLM-4.7: ~5,884 votes
• Grok 4.20-beta1: ~5,071 votes — Preliminary
• Gemini 3.1 Pro Preview: ~4,042 votes — Preliminary
• MiniMax M2.1 Preview: ~3,201 votes — Preliminary

The procurement read: Gemini 3 Pro at #5 with 39,000+ votes is, in pure statistical terms, more reliably characterized than Gemini 3.1 Pro Preview at #2 with 4,000 votes. That doesn't mean Gemini 3 Pro is "better" — it means the rank you see for it is much closer to the rank you'll see in three weeks.

Geographic and Mirror Variance

The third layer of leaderboard nuance most teams miss: the leaderboard you see depends on where you're looking from and which mirror you're reading.

U.S. Prompts vs India Prompts vs EU Prompts

LMArena aggregates votes globally, but the prompt distribution is geographically uneven. U.S. votes still dominate raw count, which biases the leaderboard slightly toward English-first prompts and U.S.-cultural-context evaluations.

• Models tuned for U.S. business English (Claude, GPT-5.2) typically rank a few Elo points higher than they would on a balanced global corpus.
• Models tuned for multilingual prompts (Gemini family, GLM-4.7, Qwen 3.5) are slightly under-rewarded on the global Text leaderboard but often over-perform in regional internal evals.
• For procurement teams in India, EU, or LATAM markets, the global LMArena rank is a starting point, not the answer. Run an internal eval on your actual prompt distribution.

OpenLM vs LMArena — Why Same Model, Different Rank

OpenLM, Hugging Face Open LLM Leaderboard, OpenRouter, and Artificial Analysis all rank LLMs — but they rank different things:

• LMArena: crowdsourced human-preference Elo on blind A/B prompts.
• Hugging Face Open LLM: automated benchmark composite (MMLU-Pro, GPQA Diamond, IFBench, MATH).
• OpenRouter: real production usage by token volume — a popularity signal, not a quality signal.
• Artificial Analysis Intelligence Index v3: weighted blend of capability + speed + cost.

A model can rank #1 on LMArena and #4 on Hugging Face. Both are correct — they're measuring different things. If your workload is conversational and human-judged, weight LMArena. If it's structured-output with verifiable correctness (math, code, factual recall), weight Hugging Face benchmarks. For an open-source-specific cross-leaderboard view, see Open-Source LMArena Rankings: 7 Models Closing the Gap.

How to Check the Live Leaderboard Without lmarena.ai

If you need programmatic access — for a procurement dashboard, a Slack bot, or an internal Confluence page — you don't have to scrape lmarena.ai. The community-maintained arena-ai-leaderboards JSON feed on GitHub mirrors the official data as structured JSON and updates within hours of LMArena's own publish cycle.

Three practical use cases:

• Daily Slack notification when a model crosses ±10 Elo from your procurement baseline.
• Procurement dashboard that pulls Elo + CI + vote count for a watchlist of 5 models.
• Internal alerting when a Preliminary tag drops on a model your team is evaluating — the moment its rank stabilizes.

For the broader month-over-month picture (and the April 2026 vote-pipeline overhaul that caused three top-10 models to lose 30+ Elo overnight), see our companion update: April 2026 LMArena Shake-Up: 3 Models Crashed Out of the Top-10.

The Bottom Line — Don't Lock Procurement to a Wednesday Screenshot

The "who is #1 right now" question has a precise answer at any given moment, but a one-week-old answer is already obsolete for any procurement-grade decision. The actionable read for enterprise buyers is not the headline rank — it's the rank-plus-CI-plus-vote-count triple read.

• Treat overlapping CIs as ties. Don't pay #1 prices for #2 capability.
• Wait for Preliminary tags to drop before signing 12-month contracts.
• Cross-reference against at least one non-LMArena leaderboard for your specific workload.
• Run an internal blind eval on your own prompts before any procurement-grade commitment.

Frequently Asked Questions (FAQ)

Sources & References

• LMArena (official) — Live LLM leaderboards including the Text top-10 baseline.
• LMArena Leaderboard Changelog — Official record of model additions and methodology changes.
• arena-ai-leaderboards JSON Feed — Open mirror of official LMArena data for programmatic access.
• Hugging Face Open LLM Leaderboard — Automated-benchmark cross-reference for verifying model performance.
• OpenRouter Rankings — Production-usage signal across LLM providers.
• Artificial Analysis — Intelligence Index v3 and capability/speed/cost blends.